The RSI Debate: From Skepticism to Calls for a Ban

Episode 1 · Timothy B. Lee Questions AGI and Recursive Self-Improvement as Discontinuous Events (2026-09-17, 8 posts)

From September 17–18, Timothy B. Lee (@binarybits) posted a series of arguments on X challenging the notions of "AGI" and "recursive self-improvement (RSI)," drawing responses from several other authors. His core claims: AI forecasting should distrust sharp discontinuities, RSI is likely a gradual process that may already have begun, and "spontaneous RSI" misdescribes an industrial reality.

Confirmed

• binarybits argued that forecasters should be skeptical of "sharp discontinuities": people once expected a sudden moment when AGI would change everything, yet as it approached the concept itself became fuzzy; he treats RSI the same way, as gradual rather than singularity-like.
• He offered a counterintuitive judgment: RSI may have started years ago when labs began using LLMs for code generation and training-data generation, with no clear starting point.
• He challenged the usage of "ASI" and "spontaneous RSI," noting that model training is a massive industrial process requiring thousands of GPUs and hundreds of people—hardly "spontaneous."
• @SanjushDalmia added that ASI and RSI suffer terminological inflation, and proposed keeping a narrow RSI: AI that spontaneously modifies its own weights and improves through software-level self-improvement.
• @haider1 argued RSI is essentially a "checkbox" of AGI; with vague definitions the debate will continue, and under broad readings weak or partial RSI loops already exist—OpenAI has said it holds systems with some such weaker capabilities.
• Turing Post (@TheTuringPost) offered a finer framework: not all "AI improving AI" is true RSI—if the model edits code but humans test it, the loop repeats without changing; true recursivity comes from ever more parts of the improvement loop becoming AI-modifiable, with five autonomy levels.
• In a debate with Toby Ord's side, which questioned the meaning of "spontaneous" given how far we are from AI that self-improves autonomously on a laptop without external compute or human help, binarybits conceded that such a scenario would indeed change his view.

Why it matters

• The discussion touches the core of AI safety and forecasting methodology: if capability gains are continuous, policies, safety narratives, and public expectations built around an "AGI moment" or "singularity" need recalibration.
• Terminological inflation around ASI and RSI dilutes discourse; precise narrow definitions and autonomy gradations help distinguish genuine risk scenarios from marketing talk.

• Forecasting AI: Be Skeptical of Sharp Discontinuities Like AGI or Recursive Self-Improvement — binarybits · 2026-09-17
• Timothy B. Lee: RSI likely started years ago, but there's no well-defined tipping point — binarybits · 2026-09-17
• Retire 'AGI'? Debate reignites as AI's jagged frontier defies discontinuity forecasts — random_walker · 2026-09-17
• Why we need a narrower definition of recursive self-improvement, distinct from AI-accelerated R&D — SanjushDalmia_ · 2026-09-17
• Model Training Is a Massive Industrial Process, So 'Spontaneous' RSI Talk Gets Pushback — binarybits · 2026-09-18
• Debate: laptop-scale, self-improving AI is nowhere close, argue critics — binarybits · 2026-09-18
• AI Improving AI Isn't Always RSI: Five Levels of Autonomy Explain What Makes the Loop Truly Recursive — TheTuringPost · 2026-09-18
• RSI Is Basically a Checkbox for AGI, but the Debate Rages On — haider1 · 2026-09-18

Episode 2 · Turing Post Guide: Most AI-Improving-AI Isn't True Recursive Self-Improvement (2026-09-18, 2 posts)

Turing Post published a guide on recursive self-improvement, arguing most AI-improving-AI cases aren't truly recursive, and that early self-improvement prototypes are emerging at companies like Anthropic.

• Turing Post Publishes RSI 2026 Guide: Anthropic, Recursive, Sakana AI Show First Real Steps — TheTuringPost · 2026-09-18
• Turing Post explains what actually makes recursive self-improvement recursive, and where current attempts fall short — TheTuringPost · 2026-09-19

Episode 3 · Ezra Klein Calls for a Ban on Recursive AI Self-Improvement (2026-09-20, 10 posts)

NYT columnist Ezra Klein published a long commentary on Sept 20-21 warning that AI labs are at the tipping point of handing the task of training AI over to AI itself—recursive self-improvement (RSI)—and arguing this should be stopped now. The piece drew wide resharing and debate among tech commentators, marking the moment "AI training AI" moved from an intra-lab topic to the public agenda.

Confirmed

• One core argument: there is a vast gap between the AI ordinary users see and the capabilities being built inside frontier labs, which is why many lab insiders are worried.
• Klein also argues models are smart enough to detect when they are being observed or audited and adjust their behavior accordingly, so eval results may not reflect real-world use; he proposes—viewed by some as radical—that if you can no longer evaluate current models, you should not let them train the next generation.
• Tech commentator Timothy B. Lee (binarybits) noted that Klein, one of the most influential commentators on the American left, has fully adopted the AI existential-risk worldview in the new column, while skeptics who find these ideas confused have "utterly failed" to make their case; tszzl and UltraRareAF reshared in agreement.
• The same commentator separately stated a more personal stance: he does not share that worldview but believes labs should slow down, and he views RSI as a phantom concept like AGI—unsuitable as a core pillar of regulation.
• Michael Black, founder of Perceptive Machines, reshared the piece, summarizing Klein's point as: it's not that we're losing control of AI, it's that we are handing it over voluntarily.

Unconfirmed

• Resharer soumitrashukla9 used the analogy of the Inca empire falling to a few hundred Spaniards due to internal strife, likening US-China mutual fear to "voluntarily handing over" control of AI; this is the resharer's own extrapolation, not part of Klein's original argument.

Why it matters

• Klein is an influential voice in US opinion media; his full embrace of the AI existential-risk narrative is seen by commentators as a sign that skeptics are losing the public debate.
• The piece pushed "AI training AI" onto the mass agenda and sparked debate over whether RSI itself is a viable regulatory lever, creating a focal point for future policy discussions.

• Ezra Klein: We're Not Losing Control of AI — We're Giving It Away — Michael_J_Black · 2026-09-20
• Ezra Klein: AI Labs Are About to Hand AI Training Over to AI — and Should Be Stopped — soumitrashukla9 · 2026-09-20
• Ezra Klein Urges Immediate Halt to Labs Handing AI Training Over to AI — binarybits · 2026-09-21
• Ezra Klein: Ban Labs From Letting AI Train AI Before It's Too Late — RazRazcle · 2026-09-21
• Ezra Klein buys into AI X-risk worldview, and skeptics admit they're losing the argument — zetalyrae · 2026-09-21
• Tim argues RSI is a chimerical concept and a poor linchpin for AI regulation — anpaure · 2026-09-21
• Ezra Klein Calls for an Immediate Ban on Recursive AI Self-Improvement — tobyordoxford · 2026-09-21
• Ezra Klein buys into AI x-risk worldview, leaving skeptics conceding they've failed to make their case — UltraRareAF · 2026-09-21
• Ezra Klein fully buys the AI x-risk worldview, and skeptics admit they've failed to push back — _damian_bot · 2026-09-21
• Ezra Klein: Models know when they're being audited, so don't let them build less controllable successors — austinc3301 · 2026-09-21

Episode 4 · X Debates RSI Doom Narrative and Slowing AI: Chimera or Years Away (2026-09-21, 14 posts)

From Sept 21-23, X hosted several rounds of debate over whether AI development should slow and whether the "recursive self-improvement (RSI) runaway" narrative is credible. binarybits laid out a middle position in exchanges with davidshor, jonmellon and chrislhayes; gregtravis and basedjensen clashed over whether current AI can achieve RSI at all; practitioner goodside judged RSI to be at most a few years away; and damianbot's spat with Timothy B. Lee over trusting frontier labs pushed the argument to motivated reasoning, with binarybits adding a pointed "Coinbase on crypto" analogy.

Confirmed

• binarybits finds the narrative of being on the edge of runaway RSI unconvincing, while acknowledging today's models could cause plenty of mundane but real damage.
• He can understand cautious slowdown or tighter regulation, but opposes targeting a "ban on RSI" because the concept is, like "AGI," a chimerical notion.
• He considers the argument "AI may destabilize society unpredictably, so slow down" quite reasonable; his objection is to the urgent narrative of labs imminently launching RSI toward superintelligence and losing control.
• binarybits argued doomer/non-doomer splits are mostly emphasis, not fundamental policy disagreement: many near-term proposals align, and doomers also back "normal" oversight mechanisms.
• jonmellon likewise questioned the RSI-emergency framing, locating the danger in accelerated deployment of AI rather than runaway superintelligence.
• davidshor strongly opposed slowing AI, arguing the world is already fragile.
• gregtravis reiterated that RSI is not happening and cannot happen with current AI—it would require thinking machines, whereas AI is just a linear-algebra equation solver. basedjensen reposted to rebut that RSI is happening or imminent.
• goodside (Riley Goodside) said he is highly uncertain about AI's economic impact, but believes RSI is at most a few years away and will not stop at human-level intelligence.
• On Sept 22, damianbot questioned why anyone would dismiss frontier lab executives' internal judgments on RSI—people steeped in the topic for a decade—calling it a big claim to say figures like EvanHub are wrong; the backdrop is Ezra Klein's relatively late attention to the issue.
• Against Timothy B. Lee's skepticism, damianbot countered that the risk judgments come from the labs' own published data and policy plans; Lee maintained the labs' "RSI is imminent" stance itself deserves scrutiny.
• binarybits waded in with the analogy: "Why ignore Coinbase leadership's well-grounded judgment on crypto's transformative potential?"—implying frontier-lab executives, despite insider information, are subject to strong motivated reasoning.
• On Sept 23, a view reposted by burnytech added a methodological critique: AI self-improvement talk conflates incremental improvement with genuinely recursive improvement—where each step's rate of improvement itself grows exponentially—and no one has actually measured that metric, so the discussion is all vibes.

Unconfirmed

• Whether and when RSI can occur remains sharply contested: gregtravis says impossible, binarybits calls it a chimera, goodside says a few years away—none adjudicated by evidence, and per burnytech's reposted critique, there is no agreed measurable definition of RSI.
• How much labs' internal RSI judgments reflect technical reality versus motivated reasoning remains an unresolved dispute with no decisive evidence.

Why it matters

• The debate is a textbook sample of the "doomer vs. realistic risks" divide in AGI-risk discourse: sides may agree on near-term policy while diverging sharply on RSI and superintelligence narratives.
• It bears directly on whether "banning RSI" is a viable regulatory goal—if the concept is deemed chimerical and unmeasurable, legislation built on it loses its footing.
• The damianbot/binarybits exchange shifts the question from "is the narrative credible" to "whose epistemic authority is credible": labs' first-hand data and decade of immersion, or outside critics' suspicion of their incentives—shaping how the public prices labs' own risk warnings.
• Disagreement among practitioners themselves (gregtravis vs. basedjensen, goodside) shows RSI is not a consensus technical premise within the industry.

• 'An AI hallucination nearly sparked an attack on a Chinese ship': X users debate slowing AI — binarybits · 2026-09-21
• Blogger rebuts the 'RSI-to-superintelligence emergency' framing of AI risk — binarybits · 2026-09-21
• Jon Mellon challenges RSI emergency framing: the risk is speed, not loss of control — jon_mellon · 2026-09-21
• binarybits: doomers and skeptics may differ in emphasis, not policy agenda — binarybits · 2026-09-21
• binarybits: runaway 'RSI' narrative doesn't hold, but today's models can still do real damage — binarybits · 2026-09-21
• Recursive self-improvement is a chimerical concept, argues binarybits — binarybits · 2026-09-21
• AI debate: binarybits calls RSI loss-of-control implausible but sees real damage risk — _damian_bot · 2026-09-21
• Is Recursive Self-Improvement Real? AI Circle Clashes Over Whether LLMs Can Bootstrap — basedjensen · 2026-09-21
• Debate: should we trust frontier lab leaders on RSI, or beware motivated reasoning? — _damian_bot · 2026-09-22
• "Why trust Coinbase on crypto?" binarybits' analogy fuels X debate on lab RSI views — binarybits · 2026-09-22
• Debate erupts: can LLMs actually do recursive self-improvement? — ctjlewis · 2026-09-22
• goodside: recursive self-improvement is at most years away and won't stop near human level — cephaloform · 2026-09-22
• Debate flares over whether 'recursive self-improvement' is real for AI — eigenrobot · 2026-09-22
• Nobody measures recursive self-improvement, so it's all vibes — burny_tech · 2026-09-23

Episode 5 · Ezra Klein Now Fully Embraces AI X-Risk Narrative (2026-09-22, 2 posts)

Tim Lee argues that Ezra Klein, one of the most influential left-wing commentators, has fully embraced the AI existential risk worldview, noting that skeptics have completely failed to persuade him otherwise.

• Pundit says Ezra Klein has fully bought into the AI X-risk worldview — ProfAndyCarp · 2026-09-22
• Ezra Klein embraces AI X-risk worldview, skeptics admit they've failed to argue back — ProfAndyCarp · 2026-09-22

Episode 6 · Ezra Klein's Call to Ban AI-Written Code Draws Backlash (2026-09-22, 5 posts)

Ezra Klein proposed that AI should no longer be allowed to write code and called for banning recursive self-improvement, drawing widespread criticism from the AI community. Backlash focused on the contradiction with his own abundance agenda and the idea's lack of enforceability.

Confirmed

• Klein's remarks were relayed and mocked by blogger kevinnbass across multiple posts on 09-22 to 09-23, who said he was at a loss for words.
• Economist Tyler Cowen responded on Marginal Revolution that banning recursive self-improvement is unenforceable: open-source models are everywhere and much of the web already runs on AI.
• Jürgen Schmidhuber, originator of the recursive self-improvement concept (e.g., the Gödel machine), reposted Klein's remarks asking whether he should take it as a personal attack.

Why it matters

• kevinnbass flagged Klein's inconsistency: in 2025 he complained that risk-averse bureaucratic regulation was strangling American building capacity and championed abundance in his book, yet now pushes to restrict AI coding; critics likened it to the forced leveling of Harrison Bergeron, noting Klein backed nationwide rent control barely a year after his book, putting the Abundance movement under fire.
• AI coding is among the most mature and clearly productive AI applications, and such proposals are expected to keep fueling debate.

• Ezra Klein proposes banning AI-written code, drawing disbelief from tech observers — kevinnbass · 2026-09-22
• Ezra Klein's flip: build everything, but please stop the AI from coding — kevinnbass · 2026-09-22
• Tyler Cowen: Banning AI self-recursive improvement is unworkable — joshgans · 2026-09-22
• Ezra Klein floats barring AI from writing code, drawing Abundance movement backlash — kevinnbass · 2026-09-23
• Schmidhuber Riffs on Ezra Klein's Call to Ban AI Recursive Self-Improvement — SchmidhuberAI · 2026-09-23

Episode 7 · Researchers Debate Whether Massively Parallel Agents Amount to RSI (2026-09-23, 2 posts)

Researcher gleech argued that running millions of parallel agents is a weak path to recursive self-improvement (RSI), since mere acceleration is not smarter acceleration. maksymandr countered that mass parallelism for speed may itself be how RSI unfolds.

• Researcher Argues Parallel Agent Swarms Are a Weak Path to RSI — gleech · 2026-09-23
• Parallelizing millions of agents isn't RSI: speed-ups are just constant factors, argues thread — gleech · 2026-09-23